Page 1; B&T14350-ERIM 329 omslag Glorie 01okt14
CLEARING BARTER EXCHANGE MARKETS 329 KRISTIAAN M. GLORIE KIDNEY EXCHANGE AND BEYOND Advanced computer assisted markets, otherwise known as smart markets, are becoming an important part of our modern society. This dissertation considers smart barter exchange Clearing Barter Exchange markets, which enable people to trade a wide range of goods: from shifts, to houses, to K.M. GLORIE - Clearing Barter Exchange Markets Exchange Barter Clearing - GLORIE K.M. kidneys. Centralized and computerized clearing is what makes these markets ‘smart’. The market clearing problem is to match demand and supply so as to maximize the gains of Markets trade. Trades, in this regard, need not be limited to pairwise swaps but may consist of trading cycles and chains involving multiple agents. Kidney Exchange and Beyond This dissertation presents several sophisticated market clearing algorithms that enable optimal clearing in large real-life barter exchange markets. With a particular focus on kidney exchanges, it shows how these algorithms can enable a significant alleviation of the present shortage of kidney donors and an improvement in health outcomes for kidney patients. State-of-the-art techniques are developed to allow the algorithms to be scalable, even when there are bounds on the number of simultaneous transactions, multiple objective criteria, and side constraints. Furthermore, innovative models and solution approaches are presented to allow market uncertainty, such as transaction failure, to be taken into account. The research presented in this dissertation contributes to the advancement of scientific
knowledge in combinatorial optimization and market design, particularly in the domains of Design & layout: B&T Ontwerp en advies (www.b-en-t.nl) Print: Haveka (www.haveka.nl) mathematical programming and market clearing, and aids the establishment and operation of smart barter exchange markets in the field of kidney exchange and beyond.
The Erasmus Research Institute of Management (ERIM) is the Research School (Onder - zoek school) in the field of management of the Erasmus University Rotterdam. The founding participants of ERIM are the Rotterdam School of Management (RSM), and the Erasmus School of Econo mics (ESE). ERIM was founded in 1999 and is officially accre dited by the Royal Netherlands Academy of Arts and Sciences (KNAW). The research under taken by ERIM is focused on the management of the firm in its environment, its intra- and interfirm relations, and its busi ness processes in their interdependent connections. The objective of ERIM is to carry out first rate research in manage ment, and to offer an ad vanced doctoral pro gramme in Research in Management. Within ERIM, over three hundred senior researchers and PhD candidates are active in the different research pro - grammes. From a variety of acade mic backgrounds and expertises, the ERIM commu nity is united in striving for excellence and working at the fore front of creating new business knowledge. Erasmus Research Institute of Management -
ERIM PhD Series Research in Management
Erasmus Research Institute of Management - Tel. +31 10 408 11 82 Rotterdam School of Management (RSM) Fax +31 10 408 96 40 M I R E Erasmus School of Economics (ESE) E-mail [email protected] Erasmus University Rotterdam (EUR) Internet www.erim.eur.nl P.O. Box 1738, 3000 DR Rotterdam, The Netherlands CLEARING BARTER EXCHANGE MARKETS Kidney Exchange and Beyond
1_Erim Glorie[stand].job 2_Erim Glorie[stand].job Clearing Barter Exchange Markets: Kidney Exchange and Beyond
Allocatie in Ruilmarkten: Nieruitwisseling en Meer
Thesis
to obtain the degree of Doctor from the Erasmus University Rotterdam by command of the rector magnificus
Prof.dr. H.A.P. Pols
and in accordance with the decision of the Doctorate Board
The public defense shall be held on
Thursday 27 November 2014 at 11:30 hrs
by
Kristiaan Michel Glorie born in Gorinchem, the Netherlands.
3_Erim Glorie[stand].job Doctoral Committee
Promotors: Prof.dr. A.P.M. Wagelmans Prof.dr. J.J. van de Klundert
Other members: Prof.dr. W. Weimar Prof.dr. A. Viana Dr. D. Manlove
Erasmus Research Institute of Management - ERIM The joint research institute of the Rotterdam School of Management (RSM) and the Erasmus School of Economics (ESE) at the Erasmus University Rotterdam Internet: http://www.erim.eur.nl
ERIM Electronic Series Portal: http://hdl.handle.net/1765/1
ERIM PhD Series in Research in Management,329 ERIM reference number: EPS-2014-329-LIS ISBN 978-90-5892-379-0 c 2014, Kristiaan Glorie
Design: B&T Ontwerp en advies www.b-en-t.nl
This publication (cover and interior) is printed by haveka.nl on recycled paper, Revive R . The ink used is produced from renewable resources and alcohol free fountain solution. Certifications for the paper and the printing production process: Recycle, EU Flower, FSC, ISO14001. More info: http://www.haveka.nl/greening
All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means electronic or mechanical, including photocopying, recording, or by any information storage and retrieval system, without permission in writing from the author.
4_Erim Glorie[stand].job Acknowledgments
Writing down the final words of this dissertation marks the end of a great journey. I would like to take this opportunity to thank the many people that have supported me throughout this journey and have made the last four years a very rewarding experience. First of all, I would like to thank my promotors, Albert Wagelmans and Joris van der Klundert. Albert, thank you for encouraging me to become a PhD candidate. You have given me the freedom to pursue my own research path while providing useful guidance and advice whenever I needed it. I appreciate your thoroughness and punctuality and have really enjoyed working and socializing together. Whether it was for research or a general question, meetings with you have always been useful and have always left me with renewed energy. Joris, it was because of our initial conversations that I decided to study kidney exchanges. You have arranged meetings with important people in the field and you have taught me to become a better operations researcher. It has been a pleasure to work together. Thank you for all the ideas and advice you have given me. Next, I would like to thank the members of my inner doctoral committee, Willem Weimar, David Manlove, and Ana Viana, for their valuable time to evaluate this dis- sertation. Willem, as a transplant surgeon and the chairman of the Dutch transplant foundation your support has been particularly valuable to me. Thank you for bringing me in touch with the transplant community and for working together on the papers that have led to Chapters 2 and 4. David, I am very happy to have met you during the Match-up workshop in Budapest. Your knowledge about matching algorithms has deeply impressed me. Thank you for your detailed comments on an early version of Chapter 3 and on this dissertation in general. Also, thank you for your hospitality when I visited you and Gregg O’Malley in Glasgow. Ana, it was very nice to have met you and your group in Porto. I am happy that it has led to a collaboration that has resulted in the paper that is the basis for Chapter 6. Thank you also for your feedback on this dissertation and for reaching out to me personally in di cult times. You are extraordinarily committed to the welfare of those around you and I am grateful to have been able to work with you.
5_Erim Glorie[stand].job vi Acknowledgments
I am also indebted to my co-authors Guanlian Xiao, Margarida Carvalho, Miguel Constantino, Paul Bouman, Marry de Klerk, Willij Zuidema, Frans Claas, and Bernadette Haase-Kromwijk. You all have strengthened my belief that the best research is performed by working together. Thank you for your valuable contributions, without which this dissertation could not have been the same. I am thankful to Itai Ashlagi for hosting me while I visited him at MIT. Itai, although the research we performed during my visit did not end up in this dissertation, my experi- ence in Cambridge has been very valuable and you have given me great advice. You have been a tremendous source of inspiration. The data used for my research has been kindly provided by the Dutch transplant foundation, the Erasmus medical center and the national reference laboratory for histo- compatibility testing in Leiden. I would like to thank Cynthia Konijn, Marry de Klerk, and Dave Roelen for their help in providing me with this data. I would further like to express my gratitude to Wilco van den Heuvel, Erwin Hans and David Stanford for serving on my doctoral committee. Remy Spliet and Harwin de Vries have been my o ce mates for the longest part of my PhD trajectory and I would like to thank them for all the fun times, interesting conver- sations and useful discussions. I am happy that they have agreed to be my paranymphs. Remy, you are one of the most social persons I know and a natural problem solver and organizer. I am happy that we have grown close as friends. Harwin, you are always able to come up with a fresh perspective on any topic we talk about. The legendary co↵ee breaks in our o ce became even better when you started bringing good co↵ee, and it has been great that we share many things beyond just research, especially because we both have become fathers. I also would like to thank Anita Vlam for the period in which we shared an o ce during my first few months at the eleventh floor. Judith Mulder and Mathijn Retel-Helmrich, although not technically my o ce mates, often made me feel like they were. I am thankful to them for all the good laughs that we have had. Willem van Jaarsveld, while also not my o ce mate, often took the liberty of rearranging or relocating objects in my o ce. Fortunately, he made up for this by his priceless insights and delicious shrimp croquettes. Twan Dollevoet has been the master of analytical puzzles and always provided helpful advice. Lanah Evers brought up some of the most lively discussions. Ev¸senKorkmaz, Ilse Louwerse, Zahra Mobini Dehkordi and Evelot Duijzer often stopped by with a warm smile and helped create a joyous environment. The sta↵ and PhD candidates at the econometric institute are easily the best part of the Erasmus to me. I am grateful to Remy, Harwin, Judith, Mathijn, Twan, Willem,
6_Erim Glorie[stand].job vii
Lanah, Ilse, Zahra, Charlie, Pieter, Evelot, Gertjan, Albert, Wilco, Adriana, G¨on¨ul, Jan, Martyn, Rommert, Dennis, Tommi, Alex, Tao-Ying, and all the others for the many inter- esting seminars, talks, co↵ee breaks, sports events, games, lunches, dinners, conferences, and other intellectual and social get-togethers. Of course, I should not forget to thank Carien, Marjon, Anneke, Ursula, and Marianne for their administrative support. We have also often been joined by Ev¸sen, Ba¸sak, Paul, Luuk, Evelien, Joris, Jelmer, and others from the Rotterdam School of Management and I would like to thank them too for four memorable years. Kevin Dalmeijer and Nico van der Windt have helped apply my research to housing markets and Jacob van der Weerd is helping me take it in yet other directions. It has been great to be able to work together with these inspirational people and I look forward to continuing our e↵orts. Ifeelprivilegedtohavemetmanygreatfriendsovertheyears.Theymakesuremylife is about much more than research and I am very grateful for that. In particular I would like to thank Jurjen, Jacob, Evert, Hein, Remy, Priscilla, Michelle, Diederik, Yun-Feng, Hui-Fang, David, Ash and Najida for all the beautiful moments we have shared. My final words of thanks go to my parents and brother for their unconditional love and support. Mom, although you are no longer with us, I hope this achievement makes you proud. My family in law has also always been there for me and I really appreciate their a↵ection. My deepest thanks go to Vˆan. The last eight years have been the happiest years of my life and it has been because of you. You have given me the most precious thing in the world, our daughter Mali.
Kristiaan Glorie Rotterdam, October 2014
7_Erim Glorie[stand].job 8_Erim Glorie[stand].job Table of Contents
Acknowledgments v
1 Introduction 1 1.1 Smart barter exchange markets ...... 1 1.2 Kidney exchange ...... 2 1.3 The clearing problem ...... 3 1.4 Incentives ...... 5 1.5 Market uncertainty ...... 6 1.6 Contributionandthesisoutline ...... 6
2 Literature review: clearing in kidney exchange 9 2.1 Introduction ...... 9 2.2 History of kidney exchange ...... 10 2.3 Transplant modalities ...... 10 2.4 Imbalance ...... 14 2.5 Allocationcriteria...... 15 2.6 Participation constraints ...... 19 2.7 Clearing algorithms ...... 20 2.8 Organization ...... 22 2.9 Outlook ...... 23
3 An e cient pricing algorithm for clearing barter exchanges with branch- and-price: enabling large-scale kidney exchange with long cycles and chains 25 3.1 Introduction ...... 25 3.2 Akidneyexchangemodel ...... 29 3.2.1 Problem definition ...... 29 3.2.2 Complexity of the clearing problem ...... 31
9_Erim Glorie[stand].job xTableofContents
3.2.3 Integer programming formulations ...... 34 3.3 Iterative solution approach ...... 38 3.3.1 Branch-and-price methodology ...... 38 3.3.2 Branch-and-bound ...... 39 3.3.3 Pricing ...... 43 3.4 Simulations ...... 47 3.4.1 Static simulation with Dutch clinical data ...... 48 3.4.2 Static simulation with US population data ...... 48 3.4.3 Sparse pools ...... 50 3.4.4 Dynamic simulations ...... 50 3.5 Computational results ...... 54 3.6 Conclusions ...... 57
4 Coordinating unspecified living kidney donation and transplantation across the blood type barrier in kidney exchange 63 4.1 Introduction ...... 63 4.2 Materials and methods ...... 66 4.2.1 Data ...... 66 4.2.2 Simulation ...... 67 4.2.3 Market clearing ...... 67 4.2.4 Policies ...... 68 4.2.5 Match failure, reneging and attrition ...... 68 4.2.6 Bridge donors ...... 69 4.2.7 Base case ...... 69 4.2.8 Sensitivity analysis ...... 70 4.2.9 Statistical analysis ...... 70 4.3 Results ...... 71 4.3.1 Nationalcoordination...... 71 4.3.2 Allowing longer chain lengths ...... 72 4.3.3 Allowing ABOi transplantation for highly sensitized patients . . . . 73 4.3.4 Sensitivity analysis population composition ...... 73 4.3.5 Sensitivity analysis unspecified and bridge donor availability (chain termination) ...... 79 4.3.6 Sensitivity analysis timing of exchanges ...... 79 4.3.7 Sensitivity analysis renege rate ...... 81 4.4 Discussion ...... 81
10_Erim Glorie[stand].job Table of Contents xi
5 Health value analysis of allocation policies in kidney exchange 85 5.1 Introduction ...... 85 5.2 An individualized health value model for kidney exchange ...... 87 5.2.1 States ...... 87 5.2.2 Transition probabilities ...... 89 5.2.3 Distribution over states ...... 94 5.2.4 Calculating QALYs gained ...... 94 5.3 Allocationpolicies ...... 95 5.3.1 Policies ...... 95 5.3.2 Determining the allocation by mixed integer programming . . . . . 96 5.3.3 Determining the maximum possible gain in health value ...... 98 5.4 Simulator ...... 99 5.4.1 Data ...... 99 5.4.2 Simulations ...... 99 5.4.3 Statistical analysis ...... 100 5.5 Results ...... 100 5.6 Conclusion and discussion ...... 105
6 Robust barter exchange 107 6.1 Introduction ...... 107 6.2 Mathematical problem description ...... 111 6.2.1 A model for market uncertainty ...... 111 6.2.2 The robust exchange problem ...... 112 6.2.3 Simple recourse ...... 113 6.2.4 Back-arcs recourse ...... 114 6.2.5 Full recourse ...... 115 6.3 Solving the robust exchange problem with simple recourse ...... 116 6.4 Solving the robust exchange problem with back-arcs recourse ...... 119 6.5 Solving the robust exchange problem with full recourse ...... 132 6.5.1 Delayed scenario generation ...... 132 6.5.2 Full recourse linearization ...... 133 6.5.3 The scenario generation subproblem ...... 135 6.5.4 Solving the scenario generation subproblem ...... 136 6.6 Refinement of the robust solution ...... 139 6.7 Computational results ...... 140 6.7.1 Instance generator ...... 140
11_Erim Glorie[stand].job xii Table of Contents
6.7.2 Simple recourse ...... 141 6.7.3 Back-arcs recourse ...... 142 6.7.4 Full recourse ...... 142 6.7.5 Protecting highly sensitized patients ...... 143 6.8 Conclusions ...... 144
7 Summary and Conclusions 151
Appendix A Estimating the probability of match failure due to positive crossmatch 155
References 163
Nederlandse Samenvatting (Summary in Dutch) 179
About the author 185
12_Erim Glorie[stand].job Chapter 1
Introduction
1.1 Smart barter exchange markets
Advanced computer assisted markets, otherwise known as smart markets, are becoming an important part of our modern society (Bichler et al., 2010). Smart markets rely on computers for operation, for ubiquitous access (e.g. through the internet), for trustworthy intermediation, and for determining market outcomes. Typically, all agents report their preferences to a centralized market operator or clearing house, and the operator or clearing house then provides an allocation and transfer prices so as to optimally match demand and supply (Roth, 2008). The problem of determining the optimal allocation and transfer prices is known as the market clearing problem. In this thesis we consider the clearing problem for barter exchange markets. Barter exchange is one of the oldest and most straightforward forms of economic activity (Smith, 1937). It concerns the direct trading of one product or service for another. Nearly everyone will, over the course of their lives, have engaged in some form of barter or another. Trivial examples include trading collectibles such as marbles or sports pictures, trading books, or perhaps, trading a shift with a colleague. Typically, these forms of barter are bilateral and involve two agents whose disposable possessions mutually suit each other’s wants. This reveals an important di culty of barter exchange: there must be a coincidence of wants (Jevons, 1875). To quote Jevons:
There may be many people wanting, and many possessing those things wanted; but to allow for an act of barter, there must be a double coincidence, which will rarely happen. ... The owner of a house may find it unsuitable, and may have his eye upon another house exactly fitted to his needs. But even if the owner of this second house wishes to part with it at all, it is exceedingly
13_Erim Glorie[stand].job 2Introduction
unlikely that he will exactly reciprocate the feeling of the first owner, and wish to barter houses. (Jevons, 1876, chapter 1)
Digital marketplaces can overcome the complexities of finding a coincidence of wants (Roth, 2008). Even opportunities for multilateral exchanges, involving many agents and many goods, can now be identified using computer algorithms. Examples of digital barter markets include house exchanges (in which agents seek to simultane- ously buy each others houses, see for example www.besthouseswap.com), service ex- changes (e.g., www.swapright.com), intra-organizational skilled worker exchanges (e.g., www.sta↵share.co.uk), and book exchanges (see for example www.readitswapit.co.uk). Arguably, the most advanced barter exchange markets operated today are kidney ex- change markets, which aim to enable transplants between incompatible patient-donor pairs (Rapaport, 1986; Roth et al., 2004; de Klerk et al., 2005).
1.2 Kidney exchange
In the United States alone, over 640,000 patients are presently su↵ering from end-stage renal disease (United States Renal Data System (USRDS), 2013). 430,000 of these pa- tients are being treated with dialysis, which means their blood has to be filtered several times a week for multiple hours. The quality of life on dialysis is low and the annual mortality rate is over 20 % (United States Renal Data System (USRDS), 2013). Kidney transplantation has been established as the preferred alternative treatment (Wolfe et al., 1999). Compared to dialysis, it o↵ers substantial advantages in terms of quality of life, patient survival, and costs (Port et al., 1993; Franke et al., 2003; Winkelmayer et al., 2002): on average, patients who receive a kidney transplant live 10 years longer than patients who remain on dialysis (Port et al., 1993), while the long term costs of trans- plantation are 4 to 5 times lower (Winkelmayer et al., 2002). Unfortunately, the number of kidneys available for transplantation is still largely insu cient to meet demand: only about 17,000 US patients can receive a transplant each year (SRTR, 2011). Kidney transplants can come from both deceased and living donors. Deceased donor kidneys are allocated to patients by means of a waiting list, which in the US currently contains 108,571 patients and has an average waiting time of 4 years (SRTR, 2011). Living donors, such as a brother or sister of the patient, can provide a direct transplant. Grafts taken from living donors generally function twice as long as grafts taken from deceased
14_Erim Glorie[stand].job 1.3 The clearing problem 3
donors (SRTR, 2011). However, in over 30 % of the cases, a living donor and his or her intended recipient are medically incompatible (Segev et al., 2005b). Kidney exchanges allow incompatible patient-donor pairs to swap donors in order to proceed with transplantation. If a patient’s donor is compatible with some other patient, and the donor of the other patient is compatible with the first patient, the patients can swap donors so that both patients can obtain a transplant (de Klerk et al., 2005; Roth et al., 2004). Due to the large potential for increasing the number of transplants, many countries have developed kidney exchange programs. Leading examples are the Netherlands, the US, the UK, Australia, and South Korea (Keizer et al. (2005), Manlove and O’Malley (2012), Park et al. (1999), Delmonico et al. (2004)). Compatibility in kidney exchange is determined by two factors: blood type compat- ibility and human leukocyte antigen (HLA) compatibility. In both cases, a patient and donor are incompatible if the patient has antibodies against an antigen contained in the donor’s cells as then the patient’s immune system will reject the donor’s tissue. There are four blood types, A, B, AB, and O, corresponding to the presence of the antigens A and B. If the donor’s blood contains an antigen that is not present in the patient’s blood, the patient will have antibodies against the donor. The HLA system contains many anti- gens, and testing for antibodies against any of the donor’s HLA is done by a so-called crossmatch test, which combines the patient’s and donor’s serum. If the crossmatch test is positive, the donor and patient are incompatible. In kidney exchange, the trading preferences of agents are directly related to the com- patibility structure. In most kidney exchange markets agents are assumed to be indi↵erent between compatible donors. However, because of the compatibility structure, some pa- tients will have a disadvantaged position. These are in particular blood type O patients and highly sensitized patients, i.e. patients with antibodies against a large part of the donor population. Patient sensitization is measured by the percentage of panel reactive antibodies (PRA), which provides an estimate of the percentage of donors with whom the patient will have a positive crossmatch test. Patient’s with a PRA of 80 % or more are considered to be highly sensitized.
1.3 The clearing problem
Given a set of agents, the objects they brought to the market, and the agent’s reported preferences over objects, the clearing problem in barter exchange markets is to determine an allocation of objects to agents so as to maximize the gains of trade. In general, there may be side payments to compensate for unequal exchanges. Every agent may for instance
15_Erim Glorie[stand].job 4Introduction
have an asking price for the good he brought to the market and a maximum buying price for every good he is interested in. As barter exchange markets are a special case of matching markets in which one side (e.g. patients) is matched to another (e.g. donors) (Demange and Gale, 1985), the clearing problem in barter exchange markets is related to the maximum matching problem (Edmonds, 1965) which is a classic combinatorial optimization problem. The fundamental aspect in clearing barter exchange markets is that if an agent’s object is allocated to another agent, the first agent should be allocated another object. Of course, there may be agents that provide goods without requiring goods in return and agents that want to obtain goods without providing any, but the actual barter takes place between agents that both provide and demand goods. Barter exchange need not be limited to pairwise exchanges but may involve trading cycles in which each participant gives an object to the next participant in the cycle and receives an object from the previous participant (Shapley and Scarf, 1974; Roth et al., 2007). Alternatively, agents that only provide a good without requiring a good in return may initiate a trading chain which ends with an allocation to an agent that does not provide any good. In practice, there typically is a constraint on the number of participants in a trading cycle or chain. For example, to avoid reneging of donors, transplants in kidney exchange cycles are typically required to be performed simultaneously and the number of logistically feasible simultaneous transplants is limited. It is precisely this constraint that makes the clearing problem in barter exchange markets substantially more di cult to solve than the classical maximum matching problem (Abraham et al., 2007). The clearing problem in barter exchange markets can also be related to the winner determination problem in combinatorial auctions (Cramton et al., 2006). The di↵erence is that in barter exchange markets agents are not assumed to have preferences over com- binatorial structures (i.e. packages) of objects, but the selected allocation must consist of combinatorial structures of agents (i.e. cycles and chains). Solving the clearing problem for barter exchange markets in an acceptable amount of time requires the aid of sophisticated algorithms and significant computing power. This is particularly the case when the market contains hundreds or even thousands of agents and there are many complex constraints or objectives. In kidney exchange, for example, the primary objective is typically to maximize the number of transplants, but there can be many secondary objectives, such as minimizing waiting times or inequity (De Klerk et al., 2010). The most successful exact algorithms in the scientific literature to solve the clearing problem are so-called ‘branch-and-price’ algorithms (Abraham et al., 2007).
16_Erim Glorie[stand].job 1.4 Incentives 5
In practice, the clearing problem is typically considered in a static or o✏ine context, in which exchanges are conducted at fixed time intervals and the assignment is optimized for the present state of the market. Of course, current agents and objects may disappear from the market and new agents may arrive. An alternative would therefore be to consider a dynamic or online context, in which the timing of exchanges is explicitly taken into account and the allocation is optimized with respect to the future evolution of the market. Such adynamiccontextmayhaveimplicationsforwhatisconsideredanoptimalallocation.
1.4 Incentives
In order to select an optimal allocation, accurate information regarding the agents, the objects, and the preferences is required. In case this information is self reported by the agents, there may be opportunities for agents to manipulate the information they reveal in order to try to achieve a better market outcome. Depending on the nature and severity of the manipulations, it may be necessary to implement incentive constraints that ensure that agents can never obtain a better outcome by providing false or incomplete information. There are two main types of incentive constraints considered in the market design literature: individual rationality constraints (Roth, 1977) and incentive compatibility constraints (Myerson, 1979). Individual rationality constraints, which are alternatively known as participation constraints, guarantee that agents or groups of agents will not be worse o↵ if they participate in the market than they would be if they did not participate in the market. Hence, under individual rationality constraints agents have no incentive to completely withhold themselves or their objects from the market. Incentive compatibility constraints, on the other hand, guarantee that no agent in the market can achieve a better market outcome by misrepresenting its information. Hence, under incentive compatibility constraints it is optimal for agents to truthfully report all their information. Given a set of agents, objects and preferences, imposing incentive constraints on the clearing problem restricts the set of possible allocations. Solving the clearing problem with these constraints may therefore result in a lower objective value compared to solving the unrestricted clearing problem. However, when using the unrestricted clearing problem the set of agents, objects and preferences used in the optimization problem may not be the full or true set of agents, objects and preferences, and this may have even more negative consequences on the objective.
17_Erim Glorie[stand].job 6Introduction
1.5 Market uncertainty
After an allocation has been selected, it may not always find continuation. In some barter exchange markets proposed transactions must be verified before they can proceed. Pro- posed transactions may fail to go forward if verification fails or if a participant withdraws. In housing markets, for example, it should be checked whether the participants in a trans- action meet the financing requirements for side payments. In kidney exchange, proposed ‘transactions’ must be checked with a final crossmatch test (in addition to the initial virtual crossmatch test) to ensure the success of eventual transplants, and patients and donors may withdraw at the last moment due to medical, psychological or other reasons (Delmonico et al., 2004; de Klerk et al., 2005; Glorie et al., 2013). In case one or more transactions fail, it may be possible to select a new allocation based on the updated information. Ideally, this new allocation is as close as possible to the initial allocation in order to minimize the material and emotional costs of the alteration. In kidney exchange markets, patients who are highly sensitized have an increased risk of match failure (besides having limited opportunities to be matched in the first place). Recovering solutions after failure may be particularly beneficial to these patients.
1.6 Contribution and thesis outline
This thesis considers the clearing problem in barter exchange markets. It focuses in particular on kidney exchange markets, but the findings are easily applicable to other types of barter exchange markets. The contributions made in this thesis are the following. In Chapter 2 we first provide an extensive literature review of the state of the art in kidney exchange clearing. In particular, we discuss the underlying principles of matching and allocation approaches, the combination of kidney exchange with other strategies such as ABO incompatible transplantation, the organization of kidney exchange, and important future challenges. Next, in Chapter 3, we consider solving the clearing problem with multiple objective criteria. We show that to achieve the best possible score on all criteria long trading cycles and chains are often needed, particularly when there are many hard-to-match patients. Long cycles and chains can be achieved by allowing some transactions to be non- simultaneous. We indicate why long cycles and chains may pose di culties for existing approaches to clearing barter exchanges. We then present a generic iterative branch-and- price algorithm which can deal e↵ectively with multiple criteria and side-constraints and
18_Erim Glorie[stand].job 1.6 Contribution and thesis outline 7
we show how the pricing problem may be solved in polynomial time in the cycle and chain length for a general class of criteria and constraints. We also study multi-center coordination of unspecified living kidney donation and ABO incompatible transplantation in kidney exchange (Chapter 4). We address questions such as whether such coordination should utilize domino paired donation (DPD) or non- simultaneous extended altruistic donor (NEAD) chains, what the length of the segments in such chains should be, when they should be terminated, what can be done to convince transplant centers to participate, and what the time interval should be between clearing rounds. To this end we integrate our aforementioned clearing algorithm with a newly developed kidney exchange simulator based on actual data from the Dutch national kidney exchange program. Chapter 5 considers the health outcomes of various allocation policies used in kidney exchange clearing. In order to analyze health outcomes, we introduce an individualized health value model for kidney exchange. This model is a Markov process with patient- donor specific transition probabilities. We also propose a new policy intended to maximize health value. This model links the individualized Markov model to the branch-and- price algorithm described in Chapter 3. We conduct long term simulations with kidney exchange data from the Netherlands. Policies are evaluated in terms of quality adjusted life years, equity, and number of transplants. Chapter 6 considers the clearing of barter exchange markets in which proposed trans- actions must be verified before they can proceed. Proposed transactions may fail to go forward if verification fails or if a participant withdraws. In case one or more matches fail, anewallocationmaybeselected.Thenewallocationshouldbeascloseaspossibleto the initial set in order to minimize the material and emotional costs of the alteration. We present a robust optimization approach that intends to maximize the number of agents selected in both the first and second allocation in a worst case scenario. Our methodol- ogy allows in particular to protect the transactions for highly-sensitized kidney exchange patients, which unfortunately are often left without a transplant using the present al- gorithms employed to clear kidney exchanges. In addition to protecting against failure, we explicitly consider the option of flexible response to failures. We do this by allowing recourse actions. We consider three recourse policies that can be easily implemented in practice. Our clearing algorithm selects an optimal planned solution taking the possibil- ity of recourse into account. If actual failures occur, our algorithm selects the optimal recourse action. Finally, in Chapter 7 we summarize and discuss the main findings of this thesis and draw some general conclusions. A Dutch summary is also provided.
19_Erim Glorie[stand].job 8Introduction
The chapters in this thesis can be read individually. Consequentially there is some overlap in the introduction to each of these chapters. The chapters are based on pa- pers that were written with various coauthors and are either published in or are (to be) submitted to scientific journals. The references to these papers are given below.
Chapter 2 K. Glorie, B. Haase-Kromwijk, J. van de Klundert, A. Wagelmans, and W. Weimar, “Allocation and matching in kidney exchange programs”. Transplant In- ternational,27(4):333-43(2014).
Chapter 3 K. Glorie, J. van de Klundert, and A. Wagelmans “Kidney exchange with long chains: an e cient pricing algorithm for clearing barter exchanges with branch- and-price”. Manufacturing & Service Operations Management,16(4):498-512(2014).1
Chapter 4 K. Glorie, M. de Klerk, A. Wagelmans, J. van de Klundert, W. Zuidema, F. Claas, and W. Weimar, “Coordinating unspecified living kidney donation and transplantation across the blood-type barrier in kidney exchange”. Transplantation, 96(9):814-20 (2013).
Chapter 5 K. Glorie, G. Xiao, and J. van de Klundert, “Health value analysis of allo- cation policies in kidney exchange”. Submitted to Operations Research (2014).
Chapter 6 K. Glorie, M. Carvalho, M. Constantino, P. Bouman, and A. Viana, “Robust barter exchange”. Working Paper (2014).
Appendix A K. Glorie, “Estimating the probability of positive crossmatch after nega- tive virtual crossmatch”. Econometric Institute report,2012-25(2012).
Summary in Dutch K. Glorie, A. Wagelmans, and J. van de Klundert, “Ethisch opti- maliseren van het ruilen van nieren”. STAtOR,13(3-4)(2012).
1A previous version of this paper has appeared as (Glorie et al., 2012b).
20_Erim Glorie[stand].job Chapter 2
Literature review: clearing in kidney exchange1
2.1 Introduction
Living kidney donation is an obvious strategy to increase the number of kidney transplants (Spital, 1988; Wolfe et al., 1999; Port et al., 1993; Franke et al., 2003; Winkelmayer et al., 2002; Lopes et al., 2013). Moreover, grafts taken from living donors generally function twice as long as grafts taken from deceased donors (SRTR, 2011). Clinical advances such as laparoscopic nephrectomy and vaginal extraction have helped increase the number of living donor kidney transplants over recent years (Segev, 2012). In the Netherlands for instance, more than half of the transplants now involves a living donor (Nederlandse Transplantatie Stichting (NTS), 2012). Nevertheless, the number of kidneys available for transplantation is still largely insu cient to meet demand: in Europe and the United States together approximately 30 patients die each day while waiting for a kidney transplant (European Society for Organ Transplantation (ESOT), 2010; United States Organ Procurement and Transplantation Network (OPTN), 2011). A major part of the problem is that, even when a living donor is willing to donate, in over 30 percent of the cases, the donor is incompatible with his or her intended recipient due to blood type or crossmatch incompatibility (Segev et al., 2005b). Several strategies have emerged to improve the utilization of living donors by mitigat- ing or overcoming the causes of incompatibility. Kidney paired donation (KPD) (Rapa- port, 1986), alternatively known as kidney exchange (Roth et al., 2004), is a strategy that allows incompatible patient-donor pairs to be matched with other incompatible pairs in
1This chapter is based on (Glorie et al., 2014b).
21_Erim Glorie[stand].job 10 Literature review: clearing in kidney exchange
order to proceed with transplantation through an exchange procedure. Other strategies include patient desensitization, living donor-deceased donor list exchange, and altruis- tic (unspecified or non-directed) donation (Montgomery et al., 2005; den Hartogh, 2010; Montgomery et al., 2006a). This review compares and discusses the various approaches to matching and allocation in kidney exchange as published in the international transplant community. In particular, it focuses on the underlying principles of market clearing and allocation approaches, the combination of KPD with other strategies such as ABO incompatible transplantation, the organization of kidney exchanges, and future challenges.
2.2 History of kidney exchange
The concept of kidney exchange was first proposed by Rapaport in 1986 (Rapaport, 1986). The initial idea was to facilitate exchange between pairs with reciprocal blood type incompatibilities (A-B and B-A), but this would later be expanded to other blood type combinations and crossmatch incompatible pairs. The first actual exchange procedure was performed in South Korea in 1991 (Kwak et al., 1999), followed by Europe in 1999 (Thiel et al., 2001), and then the US in 2000 (Zarsadias et al., 2010), the slow acceptance being mainly due to ethical and legal considerations (Ross and Woodle, 1998, 2000). After these first procedures, KPD has developed rapidly. In 2004, the Netherlands was first to launch a nationwide KPD program (de Klerk et al., 2005). Various countries have since then begun to develop national KPD programs, including the US (United Network for Organ Sharing Web Site, 2013), Australia (Ferrari et al., 2009), Canada (Canadian Blood Services Web Site, 2013), Romania (Lucan et al., 2003), and the UK (Johnson et al., 2008b,a). International exchanges, although on an ad hoc basis, have also been documented (Flanagan, 2013; La Vanguardia Ediciones, 2012).
2.3 Transplant modalities
2-Way KPD
Since the inception of KPD various transplant modalities have become available to incom- patible pairs. The simplest modality is a pairwise exchange, or 2-way KPD, between two pairs with reciprocal incompatibilities (see Figure 2.1a). In this exchange the donor of the first pair donates to the patient of the second pair, and vice versa. Usually transplants take place simultaneously so as to prevent donors from withdrawing consent after their
22_Erim Glorie[stand].job 2.3 Transplant modalities 11
Figure 2.1: Transplant modalities. Solid arrows indicate matches between donors and
recipients. Di =donori, Ri =recipienti, A =altruisticdonor,W =waitlist.
23_Erim Glorie[stand].job 12 Literature review: clearing in kidney exchange
intended recipient has received a transplant, but before they have donated themselves (Kwak et al., 1999; de Klerk et al., 2005).
k-Way KPD
Exchange can also take place between more than two pairs by generalizing the above concept to a so-called k-way KPD (Figure 2.1b). k-Way KPD which involves k pairs allows more benefits of trade to be captured as reciprocal matching is no longer required (Roth et al., 2007; De Klerk et al., 2010). In most cases, k is limited to 3 or 4 because of logistical reasons such as the simultaneous availability of operating rooms (Ferrari and de Klerk, 2009; Johnson et al., 2008b; Roth et al., 2007; De Klerk et al., 2010; Lee et al., 2009; Lucan, 2007; Montgomery et al., 2008). Although this limit is su cient to provide full benefits of trade for blood type incompatible pairs in the pool (Roth et al., 2007), highly sensitized patients could benefit if k were allowed to be larger (Ashlagi et al., 2012).
Unspecified donor chains
As an alternative to the cyclic exchange procedures described above, transplants can be organized in chain procedures. One option is to initiate a chain with an unspecified donor. Instead of donating to a patient on the deceased donor waitlist, as has been common in many countries (Johnson et al., 2008b; Gilbert et al., 2005), an unspecified donor donates to a patient of an incompatible pair (Woodle et al., 2010; Dor et al., 2011). Subsequently, the donor of that pair donates to a patient of another pair, and so forth, until the donor of the last pair in the chain donates to a patient on the deceased donor waitlist. This modality is referred to as domino-paired donation (DPD)(Figure 2.1c)(Montgomery et al., 2006a). Since it is possible to arrange the transplants in a chain such that no donor-recipient pair needs to donate a kidney before having received one, donor withdrawal can do less harm in a chain than in a k-way KPD. Therefore, the requirement of simultaneous transplants could be relaxed in chains. Non-simultaneous extended altruistic donor (NEAD) chains (Figure 2.1d) do this by recruiting ‘bridge donors who instead of donating to the deceased donor waitlist like the last donor in a DPD chain may continue the chain at a later moment in time (Rees et al., 2009). The relaxation of simultaneity allows chain procedures to involve more incompatible pairs than k-way KPD (if there is no donor withdrawal), potentially benefitting highly sensitized patients (Ashlagi et al., 2012). There has been an ongoing debate on whether it is best to use DPD or NEAD chains (Gentry et al., 2009; Ashlagi et al., 2011b; Gentry and Segev, 2011; Ireland, 2011). A recent study shows that the answer depends on the composition of the
24_Erim Glorie[stand].job 2.3 Transplant modalities 13
KPD pool and that benefits of NEAD chains are limited in case of low numbers of highly sensitized patients and su cient unspecified donors (Glorie et al., 2013).
List exchange
Another option is to initiate a chain with a list exchange (Figure 2.1e), in which the first patient in the chain does not directly receive a transplant, but instead receives priority on the deceased donor waitlist for a future deceased donor kidney, which is usually a blood type O kidney (Roth et al., 2004, 2006). The last donor of the chain again facilitates a transplant to a patient on the waitlist. However, the procedure is controversial because the latter transplant usually involves a non-blood type O kidney. Therefore list exchanges can produce disadvantages to blood type O patients on the deceased donor waitlist (den Hartogh, 2010). List exchanges have only been used in several regions in the US, where the procedure has been declared acceptable by the American Society of Transplantation (Ferrari and de Klerk, 2009).
Altruistically unbalanced exchange
All of the procedures described above can also take place with compatible pairs (Figure 2.1f). This is known as ‘altruistically unbalanced exchange donation (Ross and Woodle, 2000; Kranenburg et al., 2006). It allows incompatible pairs a better chance of finding a match, while at the same time o↵ering compatible pairs the opportunity of obtaining a better quality kidney (Gentry et al., 2007; Ratner et al., 2010; Roth et al., 2008). Studies suggest that 45 % of recipients in compatible pairs can obtain a kidney from a younger donor or a 0 mismatch kidney by participating in KPD (Gentry et al., 2007), and that approximately one third of compatible pairs would indeed be willing to do so (Kranenburg et al., 2006). Therefore altruistically unbalanced exchanges could result in both a higher number of transplants and a higher quality of transplants. Nevertheless, this form of exchange is ethically complicated as it involves asking otherwise suitable patient-donor pairs to exchange kidneys with strangers (Gentry et al., 2007).
Desensitization
Finally, there is the possibility of using desensitization techniques to overcome blood type or tissue type incompatibility. Although these techniques are costly and technically de- manding, several programs have reported promising short-term and intermediate-term results and using such techniques has become an acceptable procedure in selected indi-
25_Erim Glorie[stand].job 14 Literature review: clearing in kidney exchange
viduals (Montgomery et al., 2005; Montgomery, 2010; Montgomery et al., 2006b; Warren and Montgomery, 2010; Gloor et al., 2003; Jordan et al., 1998; Tanabe et al., 1998). In particular, desensitization for ABO incompatibility has been shown to provide good long- term graft survival, while still comparing favorably to dialysis in terms of costs (Wilpert et al., 2010; Gloor et al., 2010; Haririan et al., 2009). Combining desensitization with KPD can provide transplant opportunities to patients that would otherwise have been deemed contra-indicated and would have waited indefinitely for a suitable kidney (Claas and Dox- iadis, 2009; Montgomery et al., 2011; Sharif et al., 2012; Crew and Ratner, 2010; Glorie et al., 2014b). This is particularly true if the modalities are not just o↵ered separately to patients, but are coordinated such that hard-to-match patients can be desensitized after identifying a more favorable donor in a KPD (Glorie et al., 2013; Montgomery et al., 2011).
2.4 Imbalance
Not all incompatible patient-donor pairs have equal chances of success through KPD (Roth et al., 2007; de Klerk et al., 2011; Zenios et al., 2001; Gentry et al., 2005; Roodnat et al., 2012). This is primarily due to blood type imbalance. Because most blood type O donors can donate directly to their intended recipients, O donors will only need to enter a KPD pool if they have a positive crossmatch with their recipient. This leads to a scarcity of blood type O kidneys in KPD pools. At the same time, almost all patients are compatible with O donors. Consequentially, there will be higher demand for O kidneys than A or B kidneys, and, similarly, higher demand for A or B kidneys than AB kidneys. This leaves patient-donor pairs of types O-A, O-B, O-AB, A-AB and B-AB at a disadvantage since they need a kidney that is in higher demand than the kidney they o↵er (Roth et al., 2007). Table 1 provides a characterization of the pair types by blood type in terms of whether they are over-, under-, self-, or reciprocally-demanded (Roth et al., 2005b), and typical match results. Another imbalance is due to patient sensitization. Highly sensitized patients are at a disadvantage since they can only accept a small fraction of kidneys, mostly from donors with few HLA types, which are in high demand. Patients who are both highly sensitized and have formed an under-demanded pair will be most di cult to match. Success rates of KPD are further largely dependent on pool size and pool composition (Johnson et al., 2008b; Roth et al., 2007; Ferrari and de Klerk, 2009; Roodnat et al., 2012). The number of potential matches increases considerably with pool size. However, even in large pools typically only 50 % of pairs can be matched through KPD alone (de Klerk
26_Erim Glorie[stand].job 2.5 Allocation criteria 15
et al., 2011) (see Table 1). In the Netherlands under-demanded pairs comprise 40 % of the national pool and they have a 19 % chance of finding a match. Other pairs, which comprise 60 % of the pool, have a 73 % chance of finding a match. Because compati- ble pairs, altruistic and deceased donors typically represent the blood type frequencies of the general population, allocating these donors to KPD programs permits better match- ing. Furthermore, because blood type and tissue type distributions may di↵er between countries, international exchanges may provide benefits for selected patient-donor pairs (Flanagan, 2013; La Vanguardia Ediciones, 2012). For instance, in the Dutch KPD pro- gram 17 % of the patients have a PRA > 80 with respect to the KPD donor population, whereas in the program of the Alliance for Paired Donation in the US over 50 % of the enrolled patients have a PRA > 80 (Ashlagi et al., 2012). Part of the reason for these di↵erences may be the use of di↵erent techniques to detect unacceptable HLA specificities.
2.5 Allocation criteria
In KPD procedures, patient-donor pairs do not select the pairs with which they exchange kidneys. Instead, the allocation of donors to recipients is determined centrally. For this reason, the authority that oversees the KPD procedures must carefully consider the allo- cation criteria it will use. There can be many di↵erent perspectives as to what constitutes the best allocation. European agreements governed in the convention on human rights and biomedicine (Council of Europe, 2002) prescribe that allocation of organs should be both ‘optimal and ‘fair, without stipulating precisely what is meant by those terms. Similarly, in the United States the National Organ Transplantation Act states that donated organs should be allocated ‘equitably among transplant patients (The National Organ Transplantation Act 42, 1984). The United Network for Organ Sharing (UNOS) defines ‘equitably as a balance between utility and justice (United Network for Organ Sharing (UNOS), 1992). While ‘optimality and maximum utility is generally interpreted as achieving the maximum number of transplants, defining ‘fairness and justice is less straightforward, particularly in light of the imbalance described earlier. Although initial KPD programs have matched patient-donor pairs on an ad hoc ba- sis taking in account the above principles, most programs have now formulated precise guidelines for the allocation process (Ashlagi et al., 2011b; Keizer et al., 2005; Glorie et al., 2014d; Ferrari et al., 2011; B¨ohmig et al., 2013; Kim et al., 2007; Manlove and O’Malley, 2012). In this regard it is important to make a distinction between allocation requirements that limit the number of feasible allocations, and thereby transplants, (e.g.
27_Erim Glorie[stand].job 16 Literature review: clearing in kidney exchange eut nteDthKDprogram KPD Dutch the in results 2.1: Table hrceiaino h oiino ain-oo ye nKDposb lo yeadterhsoia match historical their and type blood by pools KPD in types patient-donor of position the of Characterization Patient blood type (% in population) B( )Oe-eaddOe-eaddOe-eaddSelf-demanded Under-demanded Over-demanded Over-demanded Self-demanded Over-demanded Under-demanded %) (3 AB Reciprocally- Over-demanded Reciprocally- %) (8 B Self-demanded Under-demanded Over-demanded Under-demanded %) (42 A Under-demanded Self-demanded %) (47 O 0 ucs ae10%scesrt ..%scesrt ..%scesrate 0%ofpool success % N.A. rate success % rate N.A. success 0%ofpool % N.A. rate success % 100 0%ofpool 1%ofpool rate rate success success % % 33 pool 100 of % rate 1 success 1%ofpool 1%ofpool rate % success 0 % 74 9%ofpool rate 1%ofpool rate success success % % 89 71 pool rate of success 4%ofpool rate % % success 6 0 8%ofpool % 67 13%ofpool rate rate success success % % 15 84 pool of % 8%ofpool rate 33 success % 20 %) (42 rate A success % 65 pool of % 16 %) (47 O demanded oo lo ye( npopulation) in (% type blood Donor demanded 8%) (8 B B( %) (3 AB
28_Erim Glorie[stand].job 2.5 Allocation criteria 17
requiring donors to be in the same age category or have the same CMV-EBV serology as their recipients) and actual allocation criteria that determine the selection of an allocation from the set of feasible allocations (e.g. maximum number of transplants between donors and recipients of the same blood type). Traditionally, deceased donor organs have been allocated to recipients in priority order. Several KPD programs have also specified a priority order for KPD allocation criteria. These include the programs operated in the Netherlands, the UK, Australia, Austria, and Korea (Keizer et al., 2005; Glorie et al., 2014d; Ferrari et al., 2011; B¨ohmig et al., 2013; Kim et al., 2007; Manlove and O’Malley, 2012). Here the criteria are hierarchical and include such factors as: maximizing the number of matched recipients, maximizing the number of blood type identical matches (to maximize the likelihood of O patients receiving akidneyandtohelpovercometheirdisadvantage),prioritizingallocationsbasedonthe number of involved recipients with a low match probability, minimizing the length of the cycles and chains, and prioritizing allocations based on waiting time of the involved recipients. Simulations show that thanks to the inclusion of the above secondary criteria, the number of highly sensitized patients matched may increase by 10% (Glorie et al., 2014d). Alternatively, criteria could also be weighted as is for example done in the UNOS KPD pilot program and the program of the Alliance for Paired Donation in the US (Ashlagi et al., 2011b; United Network for Organ Sharing (UNOS), 2012). These programs have specified weights for factors as waiting time, HLA match, PRA, prior crossmatch history, pediatric status and preferences of the incompatible pairs and their transplant centers (e.g. the distance the pair is willing to travel and whether the transplant center would accept a shipped kidney) and select the allocation that has the largest total weight (Ashlagi et al., 2011b; United Network for Organ Sharing (UNOS), 2012). Other programs have formulated requirements and criteria with regard to age, travel distance, etc. (Lucan, 2007; Kim et al., 2007; Kute et al., 2013; Ycetin et al., 2013). Two unconventional possibilities are worth mentioning. The first is the use of quality adjusted life years from transplant. Use of quality adjusted life years is commonly accepted as a prime decision criteria for many medical interventions, following the framework of Health Technology Assessment (Hutton et al., 2006; Guindo et al., 2012). However, it may conflict with commonly accepted criteria such as maximizing the number of transplants (Wolfe et al., 2008; Zenios, 2002). Another possibility is to consider long term instead of short term criteria (Unver,¨ 2009; Awasthi and Sandholm, 2009; Dickerson et al., 2012). These two do not necessarily coincide. For example, to maximize the long term number of matched patients it may be necessary to allow for some match runs in which matches
29_Erim Glorie[stand].job 18 Literature review: clearing in kidney exchange
are postponed (e.g. to allow for a future 3-way KPD to take place instead of a current 2-way KPD). After an allocation has been selected, it may not always find continuation. Proposed matches may fail due to positive crossmatch tests or patient or donor illness. In such cases a new allocation can be determined based on the updated information, as is for instance done in the Netherlands, but this requires appropriate organization of cross- matching (see Section 2.8). An alternative solution is to maintain the initial allocation as much as possible and only reallocate pairs that are part of procedures that are dis- continued. For instance, a discontinued k-way KPD could still result in multiple 2-way KPDs going forward (see Figure 2.2). The KPD program in the UK utilizes a set of hier- archical allocation criteria that aims to maximize the number of transplants that can take place after such a discontinuation (Manlove and O’Malley, 2012), by first maximizing the number of potential 2-way KPDs (including ‘back-up 2-way KPDs), and as a secondary priority maximizing the total number of transplants (Manlove and O’Malley, 2012).
Figure 2.2: Match failure. Initially a 3-way KPD is selected. If the match between donor 3 and recipient 1 fails, it is still possible to perform a 2-way KPD, either between pair 1 and pair 2, or between pair 2 and pair 3.
It can happen that di↵erent allocations rank the same on all of the selected allocation criteria. In order to select an allocation then, most programs use a deterministic tie- breaking rule (Keizer et al., 2005; Manlove and O’Malley, 2012). However, an interesting alternative for such cases is to use a stochastic rule, i.e. a lottery which selects an allocation
30_Erim Glorie[stand].job 2.6 Participation constraints 19
randomly (Roth et al., 2005b). A stochastic rule can provide several fairness properties, in particular because the probability of selecting a recipient need not be the same for all recipients and can be set in a way that alleviates the imbalance due to blood type and tissue type distributions (Roth et al., 2005b).
2.6 Participation constraints
KPD programs benefit from the participation of as many centers as possible to create a large pool. However, multi-center cooperation has brought about several di culties. Firstly, it requires consensus between participating transplant centers on the allocation requirements and criteria. Secondly, centers may judge that it is not in the interest of (some of their) patients to participate, and hence may prefer to not cooperate (fully). Thirdly, financial, scientific, or other incentives may exist, which cause cooperation to be potentially suboptimal. Thus, transplant centers may prefer to match some donors and patients locally instead of submitting them to the national pool (Glorie et al., 2013; Ashlagi and Roth, 2011a) (see Figure 2.3). One way to overcome such incentive issues is by implementing participation constraints which ensure that each transplant center can perform at least the same number of transplants in a national pool as that it can achieve on its own. Although such constraints limit the set of feasible allocations, it has been shown that they do not negatively a↵ect the long-term benefits of KPD programs (Glorie et al., 2013; Ashlagi and Roth, 2011a).
Figure 2.3: Potential participation problems. An unspecified donor (A) is registered at center A, which can generate 3 in-house transplants. In a nationally optimized program, 4 transplants are generated, but only 1 of those transplants is performed by center A.
31_Erim Glorie[stand].job 20 Literature review: clearing in kidney exchange
2.7 Clearing algorithms
Initially most KPD programs manually selected the allocation that best fit their criteria. However, given that the number of possible allocations grows exponentially with the size of the KPD pool, manually evaluating all possible allocations is only feasible for very small pools. In the US, the process of finding a match therefore originally followed a ‘first-accept scheme, which involves matching an incompatible patient-donor pair to the first pair that meets the acceptance requirements, even though matching with another pair might yield a better outcome (Segev et al., 2005a). Most KPD programs today use computer software to identify the best allocation with respect to their criteria (Ashlagi et al., 2011b; Keizer et al., 2005; Ferrari et al., 2011; B¨ohmig et al., 2013; Kim et al., 2007; Manlove and O’Malley, 2012; Kaplan et al., 2005; Hanto et al., 2008). Such software typically compares all possible allocations and can perform virtual crossmatching based on known donor HLA types and patient unacceptable HLA mismatches. However, as KPD programs expand and start to be combined with other transplant modalities, the number of possible allocations becomes so large, that even for computer programs, it becomes intractable to enumerate all feasible allocations. In these situations mathematical optimization algorithms are required to guarantee the selection of the best possible allocation (Glorie et al., 2014d; Manlove and O’Malley, 2012; Abraham et al., 2007; Constantino et al., 2013). The best current algorithms use a technique known as ‘branch-and-price’ which enables them to select an optimal allocation within minutes because they only need to consider a small subset of all possible allocations (Glorie et al., 2014d; Abraham et al., 2007). There are several aspects which provide challenges for the future. The first is that as KPD programs continue to evolve, highly sensitized and hard-to-match patients are likely to accumulate in the pool (Ashlagi et al., 2012). In such pools, the use of long chain procedures becomes essential to achieve the full benefits of exchange (Ashlagi et al., 2012). However, this renders the process of computing an optimal allocation substantially more di cult. Fortunately, recently developed algorithms have been shown to perform well even when pools are large and long chains are required (see Chapter 3). Another aspect is that taking into account the probability of match failure by maximiz- ing the expected number of transplants (which is di↵erent from maximizing the number of matches) may become more important as this will eventually lead to more transplants going forward (Pedroso, 2013; Dickerson et al., 2014; Glorie et al., 2014a). Although this still poses computational challenges, it may be an opportunity to significantly increase the success rates of KPD programs (Dickerson et al., 2013; Glorie et al., 2014a).
32_Erim Glorie[stand].job 2.7 Clearing algorithms 21
Similarly, considering online or dynamic instead of static optimization of kidney ex- changes, takes into account the timing of exchanges and the fact that patients and donors enter and leave the KPD pool over time, to optimize the desired allocation criteria in the long run (Unver,¨ 2009; Awasthi and Sandholm, 2009; Dickerson et al., 2012; Ashlagi et al., 2013). Essentially, this better represents the real decision problem underlying KPD. As of yet, because of computational complexity, it is only possible to find approximate solu- tions to the dynamic problem, but even these are often significantly better than solutions achieved through static optimization. Figure 2.4 illustrates how dynamic optimization can provide benefits. Shifting focus from static to dynamic optimization and thereby from short term to long term goals raises questions as to what defines optimality and what is equity in a dynamic setting. Full answers to these questions await further research.
Figure 2.4: Dynamic optimization. There are 5 pairs in the current KPD pool. Two 2-way KPDs are performed involving pairs 1 and 2, and pairs 3 and 4. One month later pairs 6 and 7 enter the pool. In hindsight it would now have been better to perform one 4-way KPD between pairs 1, 2, 3, and 5, and one 3-way KPD between pairs 4, 6, and 7. Dynamic optimization anticipates such situations and maximizes the expected number of transplants.
33_Erim Glorie[stand].job 22 Literature review: clearing in kidney exchange
2.8 Organization
Several countries have now implemented or pursue a national KPD program. However, there are several di↵erences in how these programs are organized. Primarily this is be- cause of geographical di↵erences: for example, the US has 244 kidney transplant centers spread out over a large area (United Network for Organ Sharing (UNOS), 2012), while the Netherlands has 8 kidney transplant centers which are relatively close together (de Klerk et al., 2005). This has immediate implications for the coordination between transplant centers and donor travel. In the Netherlands, it is feasible to move donors to the center where the matched recipient will receive the transplant. This is preferable as the recipi- ent’s home institution can provide the recipient with continuity of care and follow-up, and avoids long cold ischemic times. In the US, the retrieval surgery typically takes place at the donor center and the kidneys are shipped to the recipient’s center for transplantation. Even though this requires longer cold ischemic times and risks transportation delays, re- cent studies show comparable graft survival rates of shipped kidneys and non-shipped kidneys (Montgomery et al., 2008; Rees et al., 2009; Butt et al., 2009; Simpkins et al., 2007; Segev et al., 2011). AmajorcontributortothesuccessoftheDutchprograminestablishingconsistent high match rates, is its use of a national centralized tissues typing laboratory (Ferrari and de Klerk, 2009; de Klerk et al., 2008). In this laboratory potential donors and recipients are tested for HLA crossmatch. Having a centralized laboratory substantially enhances coordination between centers as it removes dispute about crossmatch outcomes by setting a uniform crossmatch standard. Finally, the frequency of match runs and thereby the timing of exchanges also is adi↵erentiatingelementbetweenKPDprograms.Someprogramsperformmatchruns on demand such as Korea whereas others perform them once per month, or once per three months as in the Netherlands (Glorie et al., 2013; Ferrari and de Klerk, 2009; Kim et al., 2007). Although frequent performance of match runs may result in shorter waiting time for matched recipients, it risks removing only easy-to-match pairs as the pool may not always be saturated enough for the procedures involving hard-to-match pairs to take place. Deciding when to match is therefore an important decision for KPD programs (Glorie et al., 2013). New matching software is able to advice on the optimal timing based on the composition of the pool (Unver,¨ 2009; Awasthi and Sandholm, 2009; Dickerson et al., 2012; Ashlagi et al., 2013).
34_Erim Glorie[stand].job 2.9 Outlook 23
2.9 Outlook
Since its inception in 1986, KPD has greatly expanded and has become an accepted method of transplantation at transplant centers throughout the world. Many advances have been made in terms of surgical technique, shipping donor kidneys and international exchanges. Nevertheless, many blood type O and highly sensitized patients still remain without a transplant. Important factors that have limited the success of KPD programs are logistical issues, basic trust between the various participants, and match failures. Innovative transplant modalities as altruistic donor chains and desensitization can help ameliorate the problem for critical patient groups. However, to achieve the best possible outcomes, these modalities should be coordinated jointly with KPD (Glorie et al., 2013). In this regard, this review has summarized di↵erent allocation and matching strategies. While there are many other issues that could be explored in the evolving field of KPD, matching is a key element in KPD, and by selecting the right matching strategy many patients can benefit. Future opportunities and challenges include making full use of the various modalities that are now available through integrated and optimized matching software, encouragement of transplant centers to fully participate, improving transplant rates by focusing on the expected long run number of transplants, and selecting uniform allocation criteria to facilitate international pools.
35_Erim Glorie[stand].job 36_Erim Glorie[stand].job Chapter 3
An e cient pricing algorithm for clearing barter exchanges with branch-and-price: enabling large-scale kidney exchange with long cycles and chains1
3.1 Introduction
Barter exchange markets are markets in which agents seek to directly trade their goods with each other. The trades in such markets consist of cycles in which each agent gives a good to the next agent in the cycle. Alternatively, the trades may consist of chains which are started by an agent that provides a good without requiring a good in return and end with an agent that receives a good without providing one. There are numerous examples of barter exchange markets: house exchanges (in which agents seek to simultaneously buy each others houses, see for example www.besthouseswap.com), shift exchanges (e.g., between nurses in hospitals), intra-organizational skilled worker exchanges (e.g., between projects or departments), and book exchanges (see for example www.readitswapit.co.uk). In the present paper we focus specifically on so-called kidney exchanges but our findings are easily applicable to other types of barter exchange markets.
1This chapter is based on (Glorie et al., 2014d). A previous version of it has appeared as (Glorie et al., 2012b).
37_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 26 enabling large-scale kidney exchange with long cycles and chains Kidney exchanges aim to help end-stage renal disease patients with a living and willing, but medically incompatible donor to obtain a kidney transplant, which is the preferred treatment for these patients. In particular, kidney exchanges enable patients to exchange their donor with another patient: if a patient’s donor is compatible with some other patient, and the donor of the other patient is compatible with the first patient, the patients can switch donors so that both patients can obtain a transplant (see e.g. Roth et al. (2004), de Klerk et al. (2011) and Glorie et al. (2014b)). Due to the large potential for increasing the number of transplants, many countries have developed kidney exchange programs. Leading examples are the Netherlands, the US, the UK, Australia, and South Korea (Keizer et al. (2005), Manlove and O’Malley (2012), Park et al. (1999), Delmonico et al. (2004)). Kidney exchange need not be limited to two patient-donor pairs but may involve cycles in which the donor of each pair donates, simultaneously, a kidney to the patient of the next pair in the cycle. The simultaneity is required to prevent donors from reneging after their intended recipient has received a transplant from another donor. Because of simul- taneity, the length of cycles is limited to the number of logistically feasible simultaneous transplants. Alternative to cycles, unspecified donors - i.e. donors without a specified recipient - may initiate a chain of transplants in which the last donor is allocated to the deceased donor wait list or is preserved for a future exchange. Because in a chain no patient-donor pair needs to donate before the patient in the pair has received a kidney, donor reneging in a chain would be less harmful than in a cycle. For this reason it is sometimes allowed to have one or more non-simultaneous transplants in a chain, allowing chains to be longer than cycles. Chains are increasingly common and important in clinical practice (e.g. Ashlagi et al. (2012), Glorie et al. (2013)). Presently, over 30 percent of living donors are incompatible with their intended re- cipient (Segev et al., 2005b). A patient and donor are incompatible if the donor’s blood contains an antigen that is not present in the patient’s blood, because the patient will have antibodies against such an antigen. Two cases of incompatibility can be distinguished. The first case, known as blood type incompatibility, revolves around two major antigens: A and B (blood types are denoted as AB, A, B, and O, representing the presence of these antigens). The second case, known as crossmatch incompatibility, revolves around all other antigens against which the patient may have preformed antibodies. The clearing problem in kidney exchange is to determine an assignment of donors to recipients that is feasible with respect to the medical compatibilities and maximizes one or more criteria such as the number of transplants. In practice, this problem is typically considered in a static or o✏ine context, in which exchanges are conducted at
38_Erim Glorie[stand].job 3.1 Introduction 27
fixed time intervals and the assignment is optimized for the present population, as opposed to a dynamic or online context, in which exchanges are conducted continuously and the assignment is optimized with respect to the future evolution of the population. The di culty of the clearing problem arises from the requirement that all transplants in a cycle must be performed simultaneously and that therefore cycles are limited in length. Whenever this limit is finite and larger than 2, the static clearing problem is NP-hard (Abraham et al., 2007). Abraham et al. (2007) present a mixed integer programming formulation for the clear- ing problem with the objective of maximizing a weighted sum of transplants. They solve this formulation by a branch-and-price algorithm (see Barnhart et al., 1998), in which they identify positive price variables by depth-first search. Abraham et al. (2007) show that when each transplant has equal weight in the objective function and when exchanges are limited to cycles or chains involving at most 3 patients and 3 donors, this approach works well even when the instance size is large. The main argument for limiting cycles and chains to length 3 is that initially in many pools the maximum possible number of transplants can be achieved using only cycles and chains up to length 3 (Roth et al., 2007). As we will show, however, when kidney exchange programs continue to evolve, cycles and chains up to length 3 are often not enough to attain the maximum possible number of transplants (see also Ashlagi et al. (2011b)). Moreover, when heterogenous objective weights are used or when an objective other than maximizing the sum of transplants is desired, allowing longer cycles and chains may improve the objective function. Unfortunately, with long cycles and chains depth-first pricing becomes a major bottleneck. In this paper we will show how this problem can be overcome. In practice, maximizing the (weighted) sum of transplants is not the only relevant objective criterion (see e.g. de Klerk et al. (2011)). Instead of a single weighted objective criterion, several existing kidney exchange programs use a hierarchically ordered set of criteria (e.g., De Klerk et al. (2010), Manlove and O’Malley (2012), and Kim et al. (2007)). The Dutch national kidney exchange program, in particular, uses the following hierarchical set:
Definition 3.1. Hierarchical criteria used in the Dutch kidney exchange program:
(i) Maximize the number of transplants;
(ii) Maximize the number of blood type identical transplants;
(iii) Match the patients in priority order based on ‘match probability’ (see Keizer et al., 2005);
39_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 28 enabling large-scale kidney exchange with long cycles and chains (iv) Minimize the length of the longest cycle or chain;
(v) Maximize the spread over transplant centers per cycle and chain;
(vi) Match the patient with the longest waiting time.
The Dutch criteria are based on European agreements governed in the convention on human rights and biomedicine Council of Europe (2002), which determines that the allocation of organs should be both ‘optimal’ and ‘fair’. For this reason the criteria include factors related to the probability of obtaining a transplant (criteria (ii) and (iii)) and waiting time (criterion (vi)). The exact aim of criterion (ii) is to help establish a fair allocation across patient blood types by ensuring that patients of disadvantaged blood types, such as blood type O, receive as many transplants as possible (donors of the same blood type will be reserved for them whenever this is viable). Criterion (iii) establishes such fairness in a broader sense by taking into account the total match probability (as defined in Keizer et al. (2005)). The priority order within criteria (iii) and (vi) is based on the traditional priority mechanisms for allocating deceased donor kidneys. Criteria (iv) and (v) are of a logistical nature. The hierarchy among the criteria implies that every criterion should be optimized subject to the best possible score on previous criteria. For example, the number of blood type identical transplants (criterion (ii)) should be maximized under the condition that the total number of transplants is maximum (criterion (i)).
Because of the evolution of kidney exchange pools and the ways in which exchange can take place, and because of the advent of large multi-center exchanges and the requirement of multi-criteria optimization, there is a need for new techniques for kidney exchange clearing that facilitate long chains. The work presented in this paper makes the following contributions:
1. We develop a generic iterative branch-and-price algorithm for clearing kidney ex- changes with a weighted or hierarchically ordered set of objective criteria;
2. We propose a polynomial solution method for the pricing problems as they result for a general class of criteria (which includes all criteria of the Dutch exchange);
3. The presented approach accommodates long, possibly non-simultaneous, unspecified donor chains at running times that are feasible in practice;
40_Erim Glorie[stand].job 3.2 A kidney exchange model 29
4. The approach allows for optimization for a set of transplantation centers, such as at a (inter-) national level, while taking individual rationality constraints of the participating transplantation centers into account.
This paper is organized as follows. First, Section 3.2 describes the multi-criteria kidney exchange problem mathematically. Section 3.3 details our iterative branch-and- price algorithm used to solve the multi-criteria kidney exchange clearing problem. In particular, Section 3.3.2 describes a new branching scheme and Section 3.3.3 describes how the pricing problem can be solved in polynomial time for a wide range of criteria. Section 3.4 discusses the setup of our simulations using actual kidney exchange data and Section 3.5 presents the computational results. Finally, Section 3.6 concludes the paper.
3.2 A kidney exchange model
In this section we formalize the concepts used in kidney exchanges and we mathematically define the problem under consideration.
3.2.1 Problem definition
Definition 3.2. A kidney exchange pool N consists of two sets, i.e. N = N N , where U [ S NU refers to the set of all unspecified donors and NS to the set of all incompatible specified donor-recipient pairs.
Definition 3.3. A kidney exchange graph D =(N,A) has as its node set a kidney ex- change pool N. There is an arc a =(n ,n ) A from node n N to node n N if i,j i j 2 i 2 j 2 S the donor corresponding to node ni is compatible with the recipient corresponding to node
nj.
Note that in any kidney exchange graph D =(N,A), nodes in NU ,whichcorrespond to donors without recipients, have no incoming arcs. We define a transplant cycle and a transplant chain as follows
Definition 3.4. In any given kidney exchange graph D =(N,A), a length-k cycle is an arc traversal n ,...,n such that n ,...,n N and such that (n ,n ) A and, h 1 ki { 1 k}✓ S k 1 2 for every 1 i 41_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 30 enabling large-scale kidney exchange with long cycles and chains Figure 3.1: Kidney exchange example Definition 3.5. A length k chain is an arc traversal n ,...,n such that n N and h 0 ki 0 2 U n ,...,n N and for every 0 i In practice, there exist limits on the number of transplants that can be performed simultaneously within a cycle or chain segment. This implies a natural bound on the maximum cycle and chain length. Definition 3.6. For kidney exchange graph D =(N,A) and K N, 2 C(K):= c N : c is a cycle or chain in D with length at most K ✓ n o Note that, because chains can allow for the requirement of simultaneity to be relaxed, in general the limit on chains is greater than or equal to the the limit on cycles. Typi- cally, chains use one or more non-simultaneous transplants to link several simultaneous segments of transplants. For ease of exposition, we use a single limit parameter in our no- tation. However, it is straightforward to use separate parameters and in the experiments in Section 3.5 we will do exactly this. Definition 3.7. Let D =(N,A) be a kidney exchange graph, K N, and C(K) be 2 defined as above. Then, any subset M = c1,c2,...,cM C(K), is called an exchange | | ✓ if c c = for all, 1 i, j M , i = j. i \ j ; | | 6 Thus, an exchange is a collection of interdependent kidney transplants which can be feasibly performed together. In the remainder of the paper, we assume a kidney exchange graph D =(N,A), and K N are given, and we denote with the exchange set,i.e. 2 M 42_Erim Glorie[stand].job 3.2 A kidney exchange model 31 the set of all exchanges M as defined above. Thus, an exchange set always implicitly M defines a kidney exchange graph D =(N,A), and K . Now that we have formally 2N defined exchanges, and the exchange set, we proceed by considering the criteria by which exchanges M are evaluated. 2M Definition 3.8. For any given exchange set , a criterion is a function f : R. M M! We now arrive at the formal definition of the problem under consideration: Definition 3.9. For any given exchange set and ordered set of criteria = f1,...,f , M I { |I|} a hierarchical multicriteria clearing problem is to find an exchange M ⇤ such that, 2M for each i =1,..., , M ⇤ i where i is recursively defined as i := M i 1 : |I| 2M M M { 2M fi(M) fi(M 0), M 0 i 1 with 0 := . 8 2M } M M Note that the set of criteria used in the Dutch kidney exchange program are an or- dered set of kidney exchange criteria which fit the above definition, as would be the sole criterion of maximizing the (weighted) number of transplants. Because any minimization criterion can be rewritten as a maximization criterion, the definition also accomodates for minimization criteria. Moreover, the definition also accommodates for individual ratio- nality (or participation) constraints for hospitals as sometimes required in multi-hospital settings (Glorie et al., 2013). Figure 3.1 illustrates an example of a kidney exchange clearing problem with 5 donor- recipient pairs, n1,...,n5. The bound on the length of exchange cycles K is 4. The graph has 4 feasible cycles, c = n ,n ,c = n ,n ,c = n ,n ,c = n ,n ,n ,n . 1 h 1 2i 2 h 2 3i 3 h 3 4i 4 h 1 2 3 5i There are two maximal exchanges given by M = c ,c (highlighted) and M = c . 1 { 1 3} 2 { 4} Although both exchanges have the same number of transplants, in the Dutch system exchange M1 could be preferable over exchange M2 by, for example, criterion (iv) in Definition 3.1: the maximum cycle length is 2 instead of 4. 3.2.2 Complexity of the clearing problem We will now prove that the clearing problem is NP-hard. Although this was first proved by (Abraham et al., 2007), we present here an alternative proof which is an extension of the reduction from X3C to PARTITION INTO TRIANGLES given in Garey and Johnson (1979). 43_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 32 enabling large-scale kidney exchange with long cycles and chains Consider the following decision problem: Static k-EXCHANGE (SkE) INSTANCE: A directed graph D =(N,A)andapositiveintegerk N . | | QUESTION: Is there a partition of N into disjoint sets C1,C2,...,Cm,suchthateach Ci,i=1,...,m forms a directed cycle in D of length at most k? We have the following theorem. Theorem 3.1. For 3 k N SkE is NP-Complete. | | Proof. We transform EXACT COVER BY 3-SETS (X3C) to SkE. The NP-Complete problem X3C is defined as EXACT COVER BY 3-SETS (X3C) INSTANCE: A finite set X with X =3q and a collection C of 3-element subsets of X. | | QUESTION: Does C contain an exact cover for X,thatis,asubcollectionC0 C such ✓ that every element of X occurs in exactly one element of C0? Let the set X and the collection C of 3-element subsets of X be an arbitrary instance of X3C. We shall construct a graph D =(N,A), such that a partition of N into disjoint sets C1,C2,...,Cm,witheachCi,i=1,...,m adirectedcycleinD of length at most k, exists, if and only if C contains an exact cover for X. The basic units of the X3C instance are the 3-element subsets in C.Wecanlocally replace each such subset c = x ,y,z C with the collection A of edges shown in i { i i i}2 i Figure 3.2. Then D =(N,A)isdefinedby C | | N = X a [j ]:1 j n, 1 l 9 [ { i l } i=1 [ C | | A = Ai i=1 [ Notice that the only nodes that appear in the edges belonging to more than a single Ai are those that are in the set X.IneveryAi there are n diamond like structures of k 1 nodes, where n = 12 .ItisnothardtoseethatthisinstanceofSkEcanbeconstructed in polynomial time⌃ from⌥ the X3C instance. 44_Erim Glorie[stand].job 3.2 A kidney exchange model 33 Figure 3.2: Local replacement for c =(x ,y,z) C for transforming X3C to SkE. i i i i 2 45_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 34 enabling large-scale kidney exchange with long cycles and chains If c1,c2,...,cq are the 3-element subsets from C in any exact cover for X,thenthe corresponding partition N = C C ... C of N is given by taking 1 [ 2 [ [ m a [1 ],a[1 ],x , a [1 ],a[1 ],y , a [1 ],a[1 ],z , a [n ],a[n ],a[n ] { i 1 i 2 i} { i 4 i 5 i} { i 7 i 8 i} { i 3 i 6 i 9 } ai[jl],ai[(j +1)l 1],ai[(j +1)l 2] ,j =1,...,n 1,l =3, 6, 9 from the nodes meeting A whenever c = x ,y,z is in the exact cover, and by i i { i i i} taking ai[jl],ai[jl 1],ai[jl 2] ,j =1,...,n,l =3, 6, 9 { } from the nodes meeting Ai whenever ci is not in the exact cover. This ensures that each element of X is included in exactly one 3-node cycle in the partition. Notice that cycles containing more than 3 nodes are never feasible because of the choice of the number n.Indeed,iftherewouldbeacycleC0 of length > 3, then the shortest such C0 involves adirectedpathP1 from si to ti through the gadget corresponding to some subset p and a directed path P2 from ti to si through the gadget corresponding to some subset p0,where si xi,yi,zi and ti xi,yi,zi si .ThelengthofeachofP1 and P2 is at least 2{ } 2{ }\{ } k 1 3n +1+3n,sothelengthofC0 is at least 12n +2.Becausewehavechosenn = , d 12 e it easily follows that 12n +2 k + 1 and that hence C0 is infeasible. Conversely, if N = C C ... C of N is any partition of D into cycles of 1 [ 2 [ [ m size k or less, the corresponding exact cover is given by choosing c C such that i 2 a [n ],a[n ],a[n ] = C for some j,1 j m.Weleavetothereaderthestraightfor- { i 3 i 6 i 9 } j ward task of verifying that the two partitions we have constructed are as claimed. Corollary 3.1. For 3 k N the hierarchical multicriteria clearing problem is NP- | | hard. 3.2.3 Integer programming formulations The clearing problem can be formulated as a mixed integer linear program. For example, for the single criterion of maximizing the number of transplants the clearing problem can be formulated using the so-called cycle formulation, which we describe below. Although alternative mixed integer programming formulations for the kidney exchange problem have also been investigated, the linear programming (LP) relaxation of the cycle formulation provides the strongest upper bound (Constantino et al., 2013). 46_Erim Glorie[stand].job 3.2 A kidney exchange model 35 Cycle formulation The cycle formulation, which was first presented in Abraham et al. (2007), uses a binary decision variable x for each cycle and chain c C(K)thatisdefinedas: c 2 1ifc M ⇤, xc = 2 ( 0otherwise. T Setting x = x1,...,xC(K) ,theintegerprogramisgivenby: | | ⇥ ⇤ P0: max z (x)= c x (3.1) 0 | |· c c C(K) 2X s.t. x 1 n N (3.2) c 8 2 c C(K):n c 2 X 2 x 0, 1 c C(K) c 2{ }82 In P0,theobjective(3.1)istoselectacollectionofcyclesandchainsthatmaximizes the number of transplants. The constraints (3.2) ensure that no patient or donor is contained in more than one selected cycle or chain. The number of variables in the cycle formulation can be very large (see Table 3.1 which shows the number of cycles and chains in pools based on actual data from the Dutch kidney exchange program), particularly because the number of chains grows rapidly with the number of nodes. In an exchange pool with 200 nodes there can be over a billion chains up to length 6, thus the formulation requires at least that many variables. In contrast, Abraham et al. (2007) showed that, when dealing only with cycles up to length 3, this number of variables is often not even attained in pools of 5,000 nodes or more (see Table 2 in Abraham et al. (2007)). Generalized cycle formulation The cycle formulation above can be generalized to allow for many other practically relevant criteria, including each of the criteria (i)-(vi) mentioned in the introduction. T Consider a criterion fi . As before, let x = x1,...,xC(K) denote the vector of 2I | | decision variables that indicate whether a cycle/chain c C(K)isselected.Inaddition, ⇥ 2 ⇤ for ni,mi N,letyi denote a ni 1vectorofauxiliaryvariableswhichareallowed 2 ⇥ ni C(K) ni to assume values in some subspace Fi R .Then,forwi R| |, vi R , Ai ✓ 2 2 2 47_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 36 enabling large-scale kidney exchange with long cycles and chains Nodes Arcs Cycles 4 Chains 4 Chains 6 10 50 0 1.0e+1 5.4e+1 20 192 8.00e+1 2.21e+2 1.34e+2 50 1087 2.04e+2 9.87e+3 2.51e+5 100 4443 5.07e+3 1.84e+5 2.44e+7 200 16412 1.00e+5 2.23e+6 1.34e+9 500 99501 8.58e+6 1.02e+8 5.83e+11 Table 3.1: Average number of cycles and chains over five random kidney exchange pools of the indicated size sampled from historical data of the Dutch national kidney exchange program mi C(K) mi ni mi R ⇥| |, Bi R ⇥ , and bi R ,thegeneralizedcycleformulationisgivenbythe 2 2 following integer program: Pi: T T max zi(x, yi)= wi x + vi yi (3.3) s.t. (3.2) A x + B y b (3.4) i i i i C(K) x 0, 1 | | 2{ } y F i 2 i Here, the objective (3.3) is to maximize zi(x)withrespecttofi (note that if fi were to be a minimization criterion, it can be rewritten as a maximization criterion by mul- tiplying the objective coe cients with 1). As before, the constraints (3.2) ensure that no patient or donor is contained in more than one selected cycle or chain. The general constraints (3.4) allow for various relationships between the selected cycles x and the auxiliary variables yi that are required to model fi. Ostensibly, the above formulations can also allow for multiple criteria by including a separate term in the objective function for each criterion under consideration. Each term is then multiplied with the relative weight attached to the criterion it models. As long as the weights are relatively close to each other this approach works well. However, if the criteria are hierarchically ordered, the required scaling of the weights will quickly lead to numerical instability which renders the program to be infeasible. This is, for example, the case with the six criteria used in the Dutch national program. Therefore, in the next subsection, we present a recursive formulation which models the criteria in the hierarchy without leading to numerical instability. 48_Erim Glorie[stand].job 3.2 A kidney exchange model 37 Recursive cycle formulation In this subsection we present a recursive formulation modelling hierarchical criteria that does not su↵er from numerical instability: the recursive cycle formulation.Theidea is not to capture the hierarchical multi-criteria structure into a single integer program, but instead recursively define multiple programs R1,...,R which are linked together |I| by ‘objective propagation’ constraints. This corresponds to a lexicographic optimization approach (see e.g. Isermann (1982); Rentmeesters et al. (1996)). The first program in the recursion sequence is the generalized cycle formulation of criterion f1.IncaseoftheDutchcriteria,wehaveR1 := P0,whereP0 is the program we have defined before for the maximization of the number of transplants. Then, denoting, in addition to the notation introduced above, the optimum value of R by z⇤, the programs R , i =2,..., ,arerecursivelydefinedas: i i i |I| Ri: T T max zi(x, yi)= wi x + vi yi (3.5) s.t. (3.2) A x + B y b j =1,...,i (3.6) j j j j z (x, y ) z⇤ j =1,...,i 1 (3.7) j j j C(K) x 0, 1 | | 2{ } y F j =1,...,i j 2 j As in the generalized cycle formulation, the objective (3.5) is to maximize the single criterion fi. However, the constraints (3.6) now include all the relationships required for modeling criteria f1,...,fi.Constraints(3.7)aretheobjectivepropagationconstraints, which link the program Ri to the programs R1,...,Ri 1,bypropagatingtheircorrespond- ing objective function values. The recursive cycle formulation naturally fits the definition of the hierarchical multi- criteria kidney exchange clearing problem. Indeed, the constraints (3.6) and (3.7) directly describe the sets , i =1,..., .Theexchangecorrespondingtothesolutionof Mi |I| program R is the solution to the hierarchical multi-criteria kidney exchange clearing |I| problem. 49_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 38 enabling large-scale kidney exchange with long cycles and chains T Step 0 Initialize C(K)andx = x1,...,xC(K) | | ⇥ ⇤ FOR i =1,..., DO |I| Step i Solve Ri on D: z := max C(K) f (x) i⇤ x 0,1 | | i 2{ } s.t. (3.2), f1(x) z1⇤,...,fi 1(x) zi⇤ 1 END FOR Output x := argmax C(K) f (x) ⇤ x 0,1 | | 2{ } |I| s.t. (3.2), f1(x) z1⇤,...,f 1(x) |I| z⇤ 1 |I| M ⇤ := c C(K):x⇤ =1 { 2 c } Table 3.2: Iterative algorithm for solving the hierarchical multi-criteria kidney exchange clearing problem 3.3 Iterative solution approach In this section we will develop an iterative branch-and-price algorithm for solving the hierarchical multi-criteria kidney exchange problem based on the recursive cycle formu- lation. The idea is to iteratively solve integer programs corresponding to the criteria in the hierarchy. If a program is solved, its objective function value is propagated to the integer program corresponding to the next criterion by means of an objective propaga- tion constraint. Table 3.2 gives a schematic overview of this iterative approach, where R and f (x):=f ( c C(K):x =1 )respectivelydenotetheintegerprogramcorre- i i i { 2 c } sponding to criterion fi and the objective function value of the exchange corresponding to T x = x1,...,xC(K) under criterion fi. Note that the algorithm is also valid when there | | is no⇥ hierarchy between⇤ the criteria and the criteria are captured into a single integer program. The algorithm then requires a single iteration. 3.3.1 Branch-and-price methodology Because the integer programming formulations described in Section 3.2 with one variable per cycle and chain grow exponentially in the size of the exchange pool, the (recursive) integer programs R1,...,R included in the approach of Table 3.2 are solved using |I| branch-and-price. The branch-and-price method starts with a limited subset C C(K) ✓ 50_Erim Glorie[stand].job 3.3 Iterative solution approach 39 of cycles and chains and solves the linear program (LP) relaxation of the integer program (IP) under consideration using the corresponding restricted variable set. Whenever linear programming duality conditions imply that adding variables may improve the solution value, corresponding cycles and chains are generated and added to C.Thisprocess repeats until strong duality conditions are satisfied. By doing this repeatedly for each node in a branch-and-bound tree, an optimal integral solution can be obtained. Generating columns for any of the LP relaxations to R1,...,R corresponds to gen- |I| erating cycles and chains in the kidney exchange graph. Let n denote the dual value of the constraint corresponding to node n N in (3.2). For the LP relaxation of R , 2 i i =1,..., ,wehavethefollowingreducedcostri of a cycle or chain c C(K) C: |I| c 2 \ i mi i 1 ri = w [c] A [k, c] µ w [c] ⌫ (3.8) c i n j · j,k j · j n c j=1 k=1 j=1 X2 X X X where, as before, denotes the dual value of the constraint corresponding to node n N n 2 in (3.2), µj,k denotes the dual value of the k-th constraint modeling criterion j in (3.6), and ⌫j denotes the dual value of the j-th objective propagation constraint in (3.7). In order to establish LP-optimality we search for cycles with positive reduced cost in the kidney exchange graph (see Section 3.3.3 for details on how this can be accomplished). If no such cycle can be found, the LP has been solved to optimality. If the LP solution is fractional, we branch, restricting one or more variables in the values they can assume, and then resolve the LP. At each node of the branching tree, the LP solution provides an upper bound on the restricted problem of that node. An integral lower bound can be obtained by solving the IP with the columns generated for the LP. If, at any node, the LP upper bound is no better than the best lower bound, its subtree can be pruned. If the IP lower bound matches the upper bound at the root node, the problem has been solved to optimality. 3.3.2 Branch-and-bound Branching An important and essential part of any branch-and-price procedure is the branching scheme. In the best branching scheme investigated in Abraham et al. (2007) branch- ing is performed on the cycles and chains in the kidney exchange graph. Whenever the LP solution is fractional, the cycle or chain whose corresponding variable has an LP value closest to 0.5 is selected and two branches are created, one in which the cycle’s corre- sponding variable is set to 0, and one in which it is set to 1. Branches are then explored 51_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 40 enabling large-scale kidney exchange with long cycles and chains using depth-first search. As there are are up to K N i cycles of length K or less in D, i=2 | | the branching tree may have exponential depth.P We therefore propose to branch on the arcs, of which there can be only up to N 2. We consider two branching schemes based on the following definition: Definition 3.10. An arc a A is fractional if 2 xa := xc c C(K):a c 2 X 2 is fractional. The existence of fractionally selected cycles need not immediately imply that a frac- tional arc exists. For instance, multiple fractional cycles might overlap, such that xa =1 for every arc a A. Fortunately, Theorem 3.2 establishes that this can never be true for 2 all arcs whenever the LP solution is fractional. Theorem 3.2. There exists a fractional arc if and only if the LP solution is fractional. Proof. The first implication is trivial: if a A is a fractional arc, then by definition 2 of x ,theremustbeatleastonec C(K):a c for which x is fractional. To a 2 2 c prove the other implication, suppose c1 is a fractionally selected cycle containing arcs a1,a2,...,ac1 .Ifanyarca c1 is not also covered by at least one other fractionally | | 2 selected cycle, then xa = xc1 and hence a is fractional. Therefore suppose there are one or more other fractional cycles which have at least one arc in common with cycle c1. Now, let c2 be such a fractional cycle containing, without loss of generality, arc a1 =(n1,n2)butnotarca2 =(n2,n3), and let c3,...,cm be all other fractional cycles m m containing arc a1.Therearetwooptions:either i=1 xci =1or i=1 xci < 1. In the first case, x = m x =1soarca ,andhencenoden , is totally covered, implying a1 i=1 ci 1 P 2 P that no positively valued cycle c C(K) c ,c ,...,c can cover arc a =(n ,n ), P 2 \{ 1 3 m} 2 2 3 and that therefore xa2 [xc1 , 1 xc2 ], making arc a2 fractional. In the second case, m 2 xa1 = c=1 xc < 1andthusarca1 is fractional. This completes the proof. P In our first branching scheme, we branch on groups of multiple arcs. If the LP solution is fractional, we select the node with the largest number of fractional out-arcs and then divide its out-arcs in two subsets, S1 and S2,andcreateabranchforeachsubset.Ineach branch, all the arcs of its corresponding subset are banned. The subsets S1 and S2 are 52_Erim Glorie[stand].job 3.3 Iterative solution approach 41 determined by adding arcs to S1 in non-decreasing order of xa value, until the sum of xa values of arcs in S1 is at least 0.5. The remainder of the arcs are added to S2. Theorem 3.2 guarantees us that we can always find a node with at least one fractional out-arc. In the second branching scheme, we branch on only one arc at a time. If the LP solution is fractional, we select the arc with fractional value closest to 0.5. We then create two branches: one in which the arc is banned, and one in which it is enforced. Again, Theorem 3.2 guarantees that a fractional arc always exists, and, moreover, that when we have branched on all fractional arcs, we have an integer solution. In order to enforce an arc a A in the master problem, we need to add the following 2 constraint: xc =1 (3.9) c C(K):a c 2 X 2 Adding constraint (3.9) to the master problem changes the reduced cost of a cycle or i chain. In particular, if A⇤ A is the set of enforced arcs, the reduced cost r of a cycle ✓ c or chain c C(K) C in problem R is now given by: 2 \ i i mi i 1 i rc = wi[c] n Aj[k, c] µj,k wj[c] ⌫j 1a c⇠a (3.10) · · 2 n c j=1 k=1 j=1 a A X2 X X X X2 ⇤ where, in addition to the previously introduced notation, ⇠a is the dual value of con- straint (3.9) and 1a c is an indicator function which is 1 if a c and 0 otherwise. 2 2 Note that banning an arc in the master problem is trivial, because that arc can simply be removed from the graph. Bounding In all cases, before branching, integral upper and lower bounds can be derived from the last iteration of the algorithm. For example, in Step 2 of the iterative solution algorithm, the maximum number of blood type identical transplants can not be higher than the total number of transplants determined in Step 1. Nor can it be lower than the number of blood type identical transplants in Step 1’s solution. These derived bounds are used to prune the irrelevant parts of the branching tree as soon as they violate the bounds. If the objective is to maximize the number of transplants, an upper bound can be derived by determining in polynomial time the maximum number of transplants when K = .Ifthereisalownumberofhighly-sensitizedpatients(i.e.patientswhoare 1 53_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 42 enabling large-scale kidney exchange with long cycles and chains crossmatch incompatible with many donors), Roth et al. (2007) have shown that this upper bound is tight.2 As in Abraham et al. (2007) such an upper bound can be determined by finding a maximum weight matching in a bipartite graph with donors on one side and patients on the other. Let us denote this bipartite graph as G =(U, V, E), with U denoting the patients, V denoting the donors, and E denoting the edges. Donors are connected to their own patients with a zero-weight edge and to all other compatible patients with an edge of weight 1. For each edge e E,letx be defined as: 2 e 1ife is selected, xe = ( 0otherwise. Amaximumweightmatchingcanthenbefoundinpolynomialtimebysolvingthe following LP3: max w x e · e e E X2 s.t. x =1 u U e 8 2 e= u,v E {X}2 s.t. x =1 v V e 8 2 e= u,v E {X}2 x [0, 1] e E e 2 8 2 During the branching process the initial bounds may be improved upon by the LP solutions (which provide an upper bound), or by a primal heuristic for constructing a feasible integer solution (which provides a lower bound). In all branching schemes, we use, as a primal heuristic, the solution to the IP with the columns generated for the LP. If, at any node of the branching tree, the LP upper bound is no better than the best lower bound, that node’s subtree can be pruned. If, at any node, the IP lower bound matches the upper bound at the root node, the problem has been solved to optimality. 2In their simulations Roth et al. (2007) use instances in which 10 % of the patients is highly sensitized, which in their study implies that these patients are crossmatch incompatible with 90 % of the donors. 3Although in principle a graph-based algorithm for maximum weight matching could also be used, for large instances it is often faster to solve the indicated linear program, possibly with delayed edge generation 54_Erim Glorie[stand].job 3.3 Iterative solution approach 43 3.3.3 Pricing In Abraham et al. (2007) the pricing problem is solved by traversing the kidney exchange graph D in search for a positive price cycle. In the worst case, this procedure enumerates all cycles in D and therefore is of order ( N K ), which is exponential in the size of the O | | input. In this section we present a polynomial algorithm to solve the pricing problem in (K N A )time. O | || | The algorithm requires that the reduced cost of a cycle can be expressed as a linear function of arc weights. Therefore, we first formulate the following lemma on the reduced cost of a cycle or chain in the recursive cycle formulation. Lemma 3.1. If the objective coe cients wj[c] and the constraint coe cients Aj[k, c], j =1,...,i, k =1,...,m for each cycle or chain c C(K) in problem R can be j 2 i i described as linear functions of arc weights, then there exist weights ⇡a R, for all arcs 2 a A, such that, for every cycle and chain c C(K), 2 2 i i rc = ⇡a (3.11) a c X2 i.e. the reduced cost of c can also be described as a linear function of arc weights. Proof. Let wi[c]= a c ↵i!i,a and Aj[k, c]= a c i,j!j,k,a0 for j =1,...,i and k = 2 2 1,...,mj,thenby(3.10),P P i mi i 1 i rc = wi[c] n Aj[k, c] µj,k wj[c] ⌫j 1a c⇠a · · 2 n c j=1 k=1 j=1 a A X2 X X X X2 ⇤ i mi i 1 = ↵ ! !0 µ ↵ ! ⌫ i i,a n i,j j,k,a · j,k j j,a · j a c n c j=1 k=1 a c j=1 a c X2 X2 X X X2 X X2 1a c⇠a 2 a A X2 ⇤ i mi i 1 = ↵i!i,a n ( i,j!j,k,a0 ) µj,k (↵j!j,a) ⌫j 1a A ⇠a 0 · · 2 ⇤ a= n,n c j=1 k=1 j=1 ! {X0}2 X X X i = ⇡a a= n,n c {X0}2 where 55_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 44 enabling large-scale kidney exchange with long cycles and chains i mi i 1 i ⇡a = ↵i!i,a n ( i,j!j,k,a0 ) µj,k (↵j!j,a) ⌫j 1a A ⇠a (3.12) 0 · · 2 ⇤ j=1 j=1 X Xk=1 X with n the dual value of the constraint (3.2) for node n, µj,k the dual value of the k-th constraint modeling criterion j in (3.6), ⌫j the dual value of the j-th objective propagation constraint in (3.7), and ⇠a the dual value of constraint (3.9). The linear relationship between the objective and constraint coe cients and the arcs in D holds for most criteria used in practice. In particular, all of the Dutch criteria have this property. Also, the constraints required for branching on arcs have this property. Note, however, that the constraints required to branch on cycles (as used by Abraham et al. (2007)) do not satisfy this relationship, because they require constraints to enforce the inclusion of a single cycle. Now, let us define a reversion operator as follows: Definition 3.11. For any directed cycle or chain c = n1,n2,...,nc , the directed cycle | | (respectively chain) ⌦ ↵ 1 c := n c ,nc 1,...,n1 | | | | is the reverse of c. ⌦ ↵ The pricing problems can now be solved in polynomial time through the algorithm given in Table 3.3. The algorithm first constructs the arc set A A of arcs that are not ✓ banned and then determines for each starting node n N ashortestpathuptolengthK 2 e in D =(N,A)(dependingonwhethernoden corresponds to an unspecified donor or not) using an adapted version of the Bellman-Ford method (Bellman, 1958)(Ford, 1956). For e e each node n N and k =0, 1,...,K the algorithm calculates functions f n : N R 2 k ! [{1} and gn : N N that respectively provide the weight of the shortest path between n and k ! any other node n0 N using at most k arcs, and the predecessor of node n0 N on such 2 2 ashortestn n0 path. The algorithm consists of four main steps. Before executing the main steps, Step 0transformsthearcspecificweightsobtainedfromLemma(3.1)suchthatthepricing problem becomes a minimization problem. Then, for each node n N, Step 1 initializes 2 n n n n the functions fk and gk , Step 2 calculates the function values of fk and gk in case of acycle(i.e. n N ), and Step 3 calculates the function values in case of a chain (i.e. 2 S 56_Erim Glorie[stand].job 3.3 Iterative solution approach 45 n N ). The final step, Step 4, checks whether there are cycles or chains with positive 2 U n reduced cost and, if there are, constructs them in reverse from the function values of gk . As stated in Theorem 3.3 below, the algorithm is exact, i.e. it always finds a positive price cycle or chain if one exists. In fact, for each starting node it finds the maximum weight cycle or chain of length at most K. However, it might be the case that a cycle or chain returned by the algorithm contains a subcycle (and hence is not feasible for the master problem). In the case of such a compound cycle or chain, Theorem 3.3 guarantees us that the subcycle will always have a positive price. We can choose to abort the algorithm as soon as a positive price cycle or chain is found, or it can be run to completion, possibly resulting in multiple positive price cycles and chains being identified (Nota bene, if run to completion, the algorithm will output each cycle c⇤ up to c⇤ times, therefore it | | may be desirable to filter the generated cycles for duplicates). Before providing the theorem, we first introduce the following definition: Definition 3.12. For any directed cycle c composed of simple cycles 1,..., m in D = (N,A), and arc weights ⇡i a A, the maximum simple cycle S(c) is the cycle given by a 8 2 i S(c)=argmax 1,..., m ⇡a 2{ } (a ) X2 i Theorem 3.3. C⇤ = , and, for all c⇤ C⇤, S(c⇤) C(K) and r > 0, if and only if 6 ; 2 2 S(c⇤) c C(K):ri > 0. 9 2 c Proof. Analogous to the Bellman-Ford method, we have, for each n, n0 N , k = 2 S 0,...,K,that n f (n0) = min w0 : P is an n n0 walk traversing at most k arcs k a (a P ) X2 =max w : P is an n n0 walk traversing at most k arcs a (a P ) X2 Then, obviously, n n n 1 c⇤(n):= n0,g (n0),g (g (n0)),...,n (3.13) { } =argmax ⇡i : P is an n n walk traversing at most k arcs a (a P ) X2 57_Erim Glorie[stand].job An e cient pricing algorithm for clearing barter exchanges with branch-and-price: 46 enabling large-scale kidney exchange with long cycles and chains i Step 0 Set w0 := ⇡ a A as in (3.12), C⇤ = a a 8 2 ; FOR EACH Node n N DO e 2 n n n Step 1 Set f (n):=0and, n0 N n , f (n0):= and g (n0):= 0 8 2 \{ } 0 1 0 ; Step 2 IF n N THEN 2 S set, for k =0,...,K 2, and for all n0 N, 2 n aˆ =(n00,n0):=argmina=(u,n ) A fk (u)+wa0 , 0 2 { } n n n e f (n0):=min f (n0),f (n00)+w0 , k+1 { k k aˆ} n n n n00 if fk (n00)+waˆ0 set, for k =0,...,K 1, and for all n0 N, 2 n aˆ =(n00,n0):=argmina=(u,n ) A fk (u)+wa0 , 0 2 { } n n n e f (n0):=min f (n0),f (n00)+w0 , k+1 { k k aˆ} n n n n00 if fk (n00)+waˆ0 n Step 4 For n, n0 NS,if(n0,n) A and fK 1(n0)+w0 n ,n < 0, 2 2 { 0 } 1 n ne n C⇤ C⇤ n0,gK 1(n0),gK 2(gK 1(n0)),...,n , ! [ ⌦ n ↵ and, for n N ,n0 N ,iff (n0) < 0, 2 U 2 S K 1 n n n C⇤ C⇤ n0,gK 1(n0),gK 2(gK 1(n0)),...,n ! [ ⌦ ↵ Table 3.3: Polynomial pricing algorithm 58_Erim Glorie[stand].job 3.4 Simulations 47 is a, possibly compound, maximum weight cycle with length at most K.Let 1(n),..., m(n) be the simple cycles composing c⇤(n)(ifc⇤(n) itself is a simple cycle, m = 1 and